Data
As model input, one needs satellite imagery with as much information as possible. The hyper-spectral satellite Prisma was used.
Unfortunately there is no sufficiently large, publicly available database, which relates point coordinates on a tailing to elemental concentrations, found at that coordinate. This would allow for model training in a straight forward fashion. Regarding tailings, we only had access to reduced concentrations for an entire tailing. These values however are not consistent, there is no methodology attached and it is unclear how the reduced concentrations relate to the surface material.
The found solution consists of taking databases from USGS as well as GeoROC and training our model on non-tailing point samples.
Model
As tailings are often covered by tailing ponds or vegetation, as a first step, the input pixels are filtered using the NDVI and NDWI indices.
Then a one-dimensional convolutional neural network is applied for each valid input pixel.
The returned elemental concentrations are averaged over all pixels to yield a final estimate for the entire tailing.